The Research On Cold-Start Recommendation Algorithm Based On Multi-Modal Data

People’s quality of life has been boosted by science and technology advances.At the same time,massive multi-modal data,including video,text,audio,and other data with abundant content features,has also been growing explosively.These huge data bring people the problem of information overload.And the traditional methods with the help of classification and search engines have been difficult to meet people’s different levels of needs and preferences of people,hence the recommender systems are established.Compared with the earlier recommender systems,the current recommender systems with large amounts of multi-modal data are widely studied.Furthermore,cold-start recommendation,which is one of the major problems of recommender systems,has received renewed attention from researchers due to the rich content information of multi-modal.Now,it has been gradually extended from the field of computer vision and natural language processing to the field of recommender systems with the development of deep learning technology.In particular,the emergence of graph neural network technology and related techniques,such as building graph convolutional neural networks through message passing mechanism,have a significant effect in the field of recommender systems.User-item interaction graph and attribute graph in recommender systems have been studied in different ways to improve the performance of recommender systems,but few studies have simultaneously exploited them to solve the cold-start recommendation problem.In this paper,we propose a multi-modal virtual-interaction cold-start recommendation model.At first,the user-item interaction graph and attribute graph are combined by improving the hierarchical clustering method to make full use of the abundant content features of multi-modal data.And then,we obtain the dependency graph in the form of adding virtual interactions to the user-item interaction graph.After that,graph convolution operations are performed on the dependency graph by utilizing different modal data to simulate user-item interactions.At last,the user and item representations are updated to perform cold-start recommendation.Extensive experiments are conducted on two publicly available datasets to verify the model validity,and the experimental results show that the proposed model performs well over cold-start scenarios,warm scenarios and mixed scenarios compared with the state-of-the-art models.Moreover,this paper also specifically researches each model component and multi-modal feature combination,and analyzes the validity and rationality of the model through the microscopic study.